Towards formal methods and software engineering for deep learning: Security, safety and productivity for dl systems development

Abstract

Deep Learning (DL) techniques are now widespread and being integrated into many important systems. Their classification and recognition abilities ensure their relevance for multiple application domains far beyond pure signal processing. As a machine-learning technique that relies on training instead of explicit algorithm programming they offer a high degree of productivity. But recent research has shown that they can be vulnerable to attacks and the verification of their correctness is only just emerging as a scientific and engineering possibility. Moreover DL tools are not integrated into classical software engineering so software tools to specify, modify and verify them would make them even more mainstream as software-hardware systems. This paper surveys recent work and proposes research directions and methodologies for this purpose.